on the denseness of in
TRANSCRIPT
Journal of Functional Analysis 238 (2006) 463–470
www.elsevier.com/locate/jfa
On the denseness of span{xλj (1 − x)1−λj } in C0([0,1])Tamás Erdélyi
Department of Mathematics, Texas A&M University, College Station, TX 77843, USA
Received 9 October 2005; accepted 27 January 2006
Available online 4 April 2006
Communicated by G. Pisier
Abstract
Associated with a sequence Λ = (λj )∞j=0 of distinct exponents λj ∈ [0,1], we define
H(Λ) := span{xλ0(1 − x)1−λ0 , xλ1(1 − x)1−λ1 , . . .
} ⊂ C([0,1]).
Answering a question of Giuseppe Mastroianni, we show that H(Λ) is dense in
C0[0,1] := {f ∈ C[0,1]: f (0) = f (1) = 0
}in the uniform norm on [0,1] if and only if
∞∑j=0
(1/2 − |1/2 − λj |) = ∞.
© 2006 Elsevier Inc. All rights reserved.
MSC: 41A17
Keywords: Density of function spaces; Müntz type theorems; Inequalities for Müntz polynomials
Associated with a sequence Λ = (λj )∞j=0 of distinct exponents λj ∈ [0,1], we define
Hn(Λ) := span{xλ0(1 − x)1−λ0 , xλ1(1 − x)1−λ1 , . . . , xλn(1 − x)1−λn
} ⊂ C([0,1])
E-mail address: [email protected].
0022-1236/$ – see front matter © 2006 Elsevier Inc. All rights reserved.doi:10.1016/j.jfa.2006.01.020
464 T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470
and
H(Λ) :=∞⋃
n=0
Hn(Λ) = span{xλ0(1 − x)1−λ0 , xλ1(1 − x)1−λ1 , . . .
} ⊂ C([0,1]).
In June, 2005, G. Mastroianni approached me with the following question. Let Λ = (λj )∞j=0 be
an enumeration of the rational numbers in (0,1). Is it true that H(Λ) is dense in C0([0,1]) :={f ∈ C([0,1]): f (0) = f (1) = 0} in the uniform norm on [0,1]? In this note we answer hisquestion by proving the following result.
Theorem. Let Λ = (λj )∞j=0 be a sequence of exponents λj ∈ (0,1). H(Λ) is dense in C0([0,1])
in the uniform norm on [0,1] if and only if
∞∑j=0
(1/2 − |1/2 − λj |
) = ∞. (1)
Throughout the paper we adopt the notation
‖f ‖A := supx∈A
∣∣f (x)∣∣
for complex-valued functions f defined on a set A.To prove the “if part” of the theorem, we need the lemma below. It is a well-known conse-
quence of Jensen’s formula.
Lemma 1. Suppose f is a bounded analytic function on the open disk D(a, r) centered at a
with radius r . If (zj )∞j=0 is a sequence of distinct complex numbers such that zj ∈ D(a, r) and
f (zj ) = 0 for each j = 0,1, . . . , and
∞∑j=0
(r − |zj − a|) = ∞,
then f ≡ 0 on D(a, r).
Proof of the “if part” of the theorem. Suppose H(Λ) is not dense in C0([0,1]) in the uniformnorm on [0,1]. Combining the Hahn–Banach theorem and the Riesz representation theorem, wehave a Borel measure with finite total variation on [0,1] such that
supp(μ) ∩ (0,1) �= ∅,
and
1∫xλj (1 − x)1−λj dμ(x) = 0, j = 0,1, . . . .
0
T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470 465
Then the function
f (z) :=1∫
0
tz(1 − t)1−z dμ(t)
is a bounded analytic function on the strip
S := {z ∈ C: 0 < Re(z) < 1
},
hence it is a bounded analytic function on the open disk D(1/2,1/2) centered at 1/2 with radius1/2. Since f (λj ) = 0 for each j = 0,1, . . . , and
∞∑j=0
(1/2 − |1/2 − λj |
) = ∞,
Lemma 1 implies that f ≡ 0 on D(1/2,1/2), hence by the unicity theorem f ≡ 0 on the strip S
as well. Hence
f (1/2 + iy) =1∫
0
t1/2+iy(1 − t)1/2−iy dμ(t) = 0
for every y ∈ R. Therefore, for every y ∈ R,
0 = f (1/2 + iy) =1∫
0
√t (1 − t)
(t
1 − t
)iy
dμ(t) =∞∫
−∞eixy dν(x),
with a non-zero Borel measure ν with finite total variation on R. However, this is a contradiction,see the exercise at the end of Section 2.1 of [7, Chapter VII], for instance. �Lemma 2. We have
∣∣y(1 − y)Q′(y)∣∣ �
(1 + 9
n∑j=0
λj
)‖Q‖[0,1]
for every Q ∈ Hn(Λ) and y ∈ (0,1).
Associated with a sequence Λ = (λj )∞j=0 of distinct nonnegative exponents λj , we define
Mn(Λ) := span{xλ0 , xλ1, . . . , xλn
}.
To prove Lemma 2, we need D.J. Newman’s inequality [2, Theorem 6.1.1, p. 276].
466 T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470
Lemma 3. Let b > 0. The inequality
∣∣yP ′(y)∣∣ � 9
(n∑
j=0
λj
)‖P ‖[0,b]
holds for every P ∈ Mn(Λ) and y ∈ (0, b].
Note that D.J. Newman [5] proves the above inequality with the constant 11 rather than 9.
Proof of Lemma 2. Note that every Q ∈ Hn(Λ) is of the form
Q(x) = (1 − x)P
(x
1 − x
)(2)
with some P ∈ Mn(Λ). Let y ∈ (0,1). Using Lemma 3, we obtain
∣∣y(1 − y)Q′(y)∣∣ =
∣∣∣∣y(1 − y)(1 − y)P ′(
y
1 − y
)1
(1 − y)2− y(1 − y)P
(y
1 − y
)∣∣∣∣�
∣∣∣∣yP ′(
y
1 − y
)∣∣∣∣ +∣∣∣∣(1 − y)P
(y
1 − y
)∣∣∣∣� y
1 − y
y9
(n∑
j=0
λj
)‖P ‖[0,
y1−y
] +∣∣∣∣(1 − y)P
(y
1 − y
)∣∣∣∣� 9
(n∑
j=0
λj
)(1 − y)
∥∥∥∥P
(u
1 − u
)∥∥∥∥[0,y]+ ∣∣Q(y)
∣∣
� 9
(n∑
j=0
λj
)∥∥∥∥(1 − u)P
(u
1 − u
)∥∥∥∥[0,y]+ ∣∣Q(y)
∣∣
�(
1 + 9n∑
j=0
λj
)‖Q‖[0,y] �
(1 + 9
n∑j=0
λj
)‖Q‖[0,1]. �
We need the following version of a simple lemma from [4], the short proof of which is pre-sented in this note as well.
Lemma 4. Let Γ := (γj )∞j=0 be a sequence of distinct nonnegative real numbers with
η :=∞∑
j=0
γj < ∞.
Then
∣∣P(z)∣∣ � exp
(9η‖ log z‖K
)‖P ‖[0,1], z ∈ K,
T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470 467
for every P ∈ M(Γ ) := span{xγ0, xγ1 , . . .} and for every compact K ⊂ C \ {0}.
In fact, in the “only if part” of the proof of the theorem, the following consequence ofLemma 4 is needed.
Lemma 5. Let Γ := (γj )∞j=0 be a sequence of distinct nonnegative real numbers with
η :=∞∑
j=0
γj � 1
20.
Then
‖Q‖[3/4,1] < ‖Q‖[0,1/2]
for every Q ∈ H(Γ ) = span{xγ0(1 − x)1−γ0 , xγ1(1 − x)1−γ1 , . . .}.
Proof of Lemma 4. The lemma is a consequence of D.J. Newman’s Markov-type inequality.Repeated applications of Lemma 3 with b := 1 and the substitution x = e−t imply that
∥∥(P
(e−t
))(m)∥∥[0,∞)� (9η)m
∥∥P(e−t
)∥∥[0,∞), m = 1,2, . . . ,
and, in particular
∣∣(P (e−t
))(m)(0)
∣∣ � (9η)m∥∥P(e−t )
∥∥[0,∞), m = 1,2, . . . ,
for every P ∈ M(Γ ). By using the Taylor series expansion of P(e−t ) around 0, we obtain that∣∣P(z)∣∣ � c1(K,η)‖P ‖[0,1], z ∈ K,
for every P ∈ M(Γ ) = span{xγ0 , xγ1, . . .} and for every compact K ⊂ C \ {0}, where
c1(K,η) :=∞∑
m=0
(9η)m‖log z‖mK
m! = exp(9η‖log z‖K
),
and the result of the lemma follows. �Proof of Lemma 5. This follows from Lemma 4 and the fact that every
Q ∈ Hn(Γ ) = span{xγ0(1 − x)1−γ0 , xγ1(1 − x)1−γ1 , . . . , xγn(1 − x)1−γn
}(3)
is of the form (2) with a
P ∈ Mn(Γ ) = span{xγ0, xγ1 , . . . , xγn
}. � (4)
Let Γ := (γj )∞j=0 be a sequence of distinct nonnegative real numbers with γ0 := 0. One of the
most basic properties of a Müntz space Mn(Γ ) defined in (4) is the fact that it is a Chebyshev
468 T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470
space of dimension n+1 on every A ⊂ [0,∞) containing at least n+1 points. That is, Mn(Γ ) ⊂C(A) and every P ∈ Mn(Γ ) having at least n + 1 (distinct) zeros in A is identically 0. Since anyQ ∈ Hn(Γ ) defined in (3) is of the form (2) with a P ∈ Mn(Γ ) defined in (4), the space Hn(Γ ) isalso a Chebyshev space of dimension n + 1 on every A ⊂ [0,1) containing at least n + 1 points.The following properties of the space Hn(Γ ), as a Chebyshev space of dimension n+ 1 on everyA ⊂ [0,1) containing at least n + 1 points, are well known (see, for example, [2,3,6]).
Lemma 6 (Existence of Chebyshev polynomials). Let A be a compact subset of [0,1) containingat least n + 1 points. Then there exists a unique (extended) Chebyshev polynomial
Tn := Tn{γ0, γ1, . . . , γn;A}
for Hn(Γ ) on A defined by
Tn(x) = c
(xγn(1 − x)1−γn −
n−1∑j=0
ajxγj (1 − x)1−γj
),
where the numbers a0, a1, . . . , an−1 ∈ R are chosen to minimize
∥∥∥∥∥xγn(1 − x)1−γn −n−1∑j=0
ajxγj (1 − x)1−γj
∥∥∥∥∥A
and where c ∈ R is a normalization constant chosen so that
‖Tn‖A = 1
and the sign of c is determined by
Tn(maxA) > 0.
Lemma 7 (Alternation characterization). The Chebyshev polynomial
Tn := Tn{γ0, γ1, . . . , γn;A} ∈ Hn(Γ )
is uniquely characterized by the existence of an alternation set
{x0 < x1 < · · · < xn} ⊂ A
for which
Tn(xj ) = (−1)n−j = (−1)n−j‖Tn‖A, j = 0,1, . . . , n.
Note that the existence of a unique (extended) Chebyshev polynomial with the above alterna-tion characterization can be guaranteed even if we allow that A = [0,1] rather than A ⊂ [0,1).This can be seen by a standard limiting argument. Namely we take the Chebyshev polynomialsTn,δ on the interval [0,1 − δ] first and then we let δ > 0 tend to 0.
T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470 469
Now we are ready to prove the “only if part” of the theorem. It turns out that the approachused in the corresponding part of the proof in [1] can be followed here, but, while the proof isstill rather short, the details are slightly more subtle.
Proof of the “only if part” of the theorem. Suppose now that Λ is a sequence of distinctexponents λj ∈ (0,1) such that (1) does not hold, that is,
∞∑j=0
(1/2 − |1/2 − λj |
)< ∞.
Then there are sequences Γ := (γj )∞j=0 of distinct numbers γj ∈ [0,1/4) and Δ := (δj )
∞j=0
of distinct numbers δj ∈ [0,1/4), and a finite set {α1, α2, . . . , αm} of distinct numbers αj ∈[1/4,3/4] such that
∞∑j=0
γj <1
20,
∞∑j=0
δj <1
20,
and
{λj : j = 0,1, . . .} ⊂ {γj : j = 0,1, . . .} ∪ {1 − δj : j = 0,1, . . .} ∪ {αj : j = 1,2, . . . ,m}.
Without loss of generality we may assume that γ0 := 0 and δ0 := 0. For notational convenience,let
Tn,γ := Tn
{γ0, γ1, . . . , γn; [0,1]},
Tn,δ := Tn
{1 − δ0,1 − δ1, . . . ,1 − δn; [0,1]},
T2n+m+1,γ,δ,α := T2n+m+1{γ0, . . . , γn,1 − δ0, . . . ,1 − δn,α1, . . . , αm; [0,1]}.
It follows from Lemmas 2, 5, and 7, and the Mean Value theorem that for every ε > 0 there existsa k1(ε) ∈ N depending only on (γj )
∞j=0 and ε (and not on n) so that Tn,γ has at most k1(ε) zeros
in [ε,1] and at least n − k1(ε) zeros in (0, ε). Similarly, applying Lemmas 2, 5, and 7, and theMean Value theorem to Sn,δ(x) := Tn,δ(1 − x) gives that for every ε > 0 there exists a k2(ε) ∈ N
depending only on (δj )∞j=0 and ε (and not on n) so that Tn,δ has at most k2(ε) zeros in [0,1 − ε]
and at least n − k2(ε) zeros (1 − ε,1).Now, counting the zeros of Tn,γ − T2n+m+1,γ,δ,α and Tn,δ − T2n+m+1,γ,δ,α , we can deduce
that T2n+m+1,γ,δ,α has at least n− k1(ε)− 3 zeros in (0, ε) and it has at least n− k2(ε)− 3 zerosin (1− ε,1) (we count every zero without sign change twice). Hence, for every ε > 0 there existsa k(ε) ∈ N depending only on (λj )
∞j=0 and ε (and not on n) so that T2n+m+1,γ,δ,α has at most
k(ε) zeros in [ε,1 − ε].Let ε := 1/4 and k := k(1/4). Pick k + m + 5 points
1< η0 < η1 < · · · < ηk+m+4 <
3
4 4
470 T. Erdélyi / Journal of Functional Analysis 238 (2006) 463–470
and a function f ∈ C0([0,1]) so that f (x) = 0 for all x ∈ [0,1/4] ∪ [3/4,1], while
f (ηj ) := 2(−1)j , j = 0,1, . . . , k + m + 4.
Assume that there exists a Q ∈ H(Λ) so that
‖f − Q‖[0,1] < 1.
Then Q − T2n+m+1,γ,δ,α has at least 2n + m + 2 zeros in (0,1). However, for sufficiently largen, Q − T2n+m+1,γ,δ,α is in the linear span of the 2n + m + 2 functions
xγj (1 − x)1−γj , j = 0,1, . . . , n,
x1−δj (1 − x)δj , j = 0,1, . . . , n,
and
xαj (1 − x)1−αj , j = 1,2, . . . ,m,
so it can have at most 2n + m + 1 zeros in (0,1). This contradiction shows that H(Λ) is notdense in C0([0,1]). �References
[1] P.B. Borwein, T. Erdélyi, The full Müntz theorem in C[0,1] and L1[0,1], J. London Math. Soc. 54 (1996) 102–110.[2] P.B. Borwein, T. Erdélyi, Polynomials and Polynomials Inequalities, Springer-Verlag, New York, 1995.[3] E.W. Cheney, Introduction to Approximation Theory, McGraw–Hill, New York, 1966.[4] T. Erdélyi, W. Johnson, The “Full Müntz Theorem” in Lp[0,1] for 0 < p < ∞, J. Anal. Math. 84 (2001) 145–172.[5] D.J. Newman, Derivative bounds for Müntz polynomials, J. Approx. Theory 18 (1976) 360–362.[6] T.J. Rivlin, Chebyshev Polynomials, second ed., Wiley, New York, 1990.[7] B. Sz.-Nagy, Introduction to Real Functions and Orthogonal Expansions, Oxford Univ. Press, New York, 1965.